Improving EMG Signal Classification with Transfer Learning Under Low-Data and Cross-Subject Conditions

Abstract

Surface electromyography is a non-invasive method used for monitoring muscle activity and is widely applied in rehabilitation, prosthetics, assistive robotics, and human–computer interaction. However, its practical use often remains limited by large differences between individuals and the effort required to train models for each new user. This study explores whether transfer learning can help address these challenges when using deep learning to classify hand and wrist gestures. The experiments use a dataset that includes eleven gestures, each repeated eight times by 22 healthy participants. Three training approaches are evaluated: (i) training and testing on the same subject (intra-subject), (ii) training on some subjects and testing on a new one (inter-subject), and (iii) transfer learning with and without resetting the fully connected output layer of the convolutional neural network. All models are evaluated using a leave-one-out cross-validation strategy across both subjects and repetitions. 

Results show that both transfer learning methods outperform the other two approaches in terms of classification accuracy. The best performance is observed when the fully connected layer is reset before fine-tuning (F1-score = 0.907, σ = 0.074). Wilcoxon signed-rank statistical tests confirm that these improvements are statistically significant, even when only a few repetitions are used for fine-tuning. In fact, using transfer learning with just four repetitions instead of eight achieves accuracy comparable to training from scratch on all eight repetitions. 

These findings suggest that fine-tuning pre-trained models can significantly reduce the effort needed to adapt EMG-based systems to new users, providing a practical and effective approach for developing user-friendly interfaces suited to assistive and rehabilitation applications.